In a recent exchange with the OSU folks, it was discovered that there is a gap in the MPICH test suite in that the code that tests for the correctness of handling long messages forgets to test that count is handled correctly. This led to the escape of a bug in the MVAPICH code, since corrected. Hari tracked this down and has sent a patch. Could the MPICH team look at this? Thanks, Rusty Begin forwarded message:
From: Hari Subramoni <[email protected]> Subject: Re: [lcrc-support #279383] Re: Long messages Date: October 26, 2015 at 12:19:42 PM CDT To: Rusty Lusk <[email protected]> Cc: "Low, John J." <[email protected]>, Steve Pieper <[email protected]>, "Wiringa, Robert B." <[email protected]>, "D.K. Panda" <[email protected]>, "[email protected]" <[email protected]>, "Balaji, Pavan" <[email protected]>
Hello All,
I was investigating to see how this was missed earlier. In our MPICH large-message test (test/mpi/pt2pt/large_message.c) which should've caught this, the code which tests the result of MPI_Get_count was missing.
I was checking the large_message.c test that is shipped with MPICH-3.1.4 (latest stable) and MPICH-3.2rc1 (preview release) available from the MPICH site. Both have the same issue. The large-message-test.patch should fix this.
If the patch looks good, it would be good if this gets applied to MPICH itself so that all MPICH derivatives can benefit from it.
Best Regards, Hari.